[mcp-analysis] MCP Structural Analysis - 2026-08-25

[github-actions[bot]]

First-day baseline of a rolling 30-day structural analysis of GitHub MCP tool responses across representative toolsets. 10 tools tested, 8,235 tokens consumed, average usefulness 3.6/5. Two tools tied for best (list_workflows via the mcpscripts wrapper, and the injected <github-context> block), both scoring 5/5 for clean, low-noise, immediately actionable data. The single biggest structural finding: nearly every raw GitHub MCP tool response repeats a large base64-encoded server-icon block in _meta.serverInfo on every call, inflating token cost independent of actual payload size — this overhead disappears entirely when going through the mcpscripts wrapper.

Note on tooling: pandas/matplotlib could not be installed (no network egress for pip in this sandbox), so charts below were generated with a lightweight stdlib-only SVG renderer instead of the standard visualization stack. This is day 1 of data collection, so trend charts show a single data point — trends will become meaningful after a few more runs.

Full Structural Analysis Report

Executive Summary

Metric	Value
Tools Analyzed	10
Total Tokens (Today)	8,235
Average Usefulness Rating	3.6/5
Best Rated Tools	list_workflows (actions, via mcpscripts): 5/5; workflow_context (injected): 5/5
Worst Rated Tools (tied)	get_file_contents, list_issues, list_discussions, get_label, search_code, search_users: 3/5

Usefulness Ratings for Agentic Work

Tool	Toolset	Rating	Assessment
workflow_context	workflow_context	⭐⭐⭐⭐⭐	Excellent — zero-cost injected identity metadata, no MCP call needed
list_workflows (mcpscripts)	actions	⭐⭐⭐⭐⭐	Excellent — clean schema, explicit pagination, no icon bloat
list_pull_requests	pull_requests	⭐⭐⭐⭐	Good — flat, minimal, directly actionable
list_code_scanning_alerts	code_security	⭐⭐⭐⭐	Good — unambiguous even when empty
get_file_contents	repos	⭐⭐⭐	Adequate — correct content but bloated by auto-generated README section
list_issues	issues	⭐⭐⭐	Adequate — good schema, but an integrity-policy filter silently redacted a result
list_discussions	discussions	⭐⭐⭐	Adequate — useful nested fields, but icon _meta roughly triples token cost
get_label	labels	⭐⭐⭐	Adequate — tiny payload, poor efficiency ratio due to fixed overhead
search_code	search	⭐⭐⭐	Adequate — locates files but no snippet, needs follow-up read
search_users	users	⭐⭐⭐	Adequate — minimal identity fields only, needs follow-up for profile detail

Schema Analysis

Tool	Type	Depth	Key Fields	Notes
workflow_context	markdown	1	actor, repository, workspace, workflow-run-id	Prompt-injected, not an MCP call
get_file_contents	object	3	content.text, resultType, _meta.serverInfo	Icon metadata + raw file text
list_issues	array	2	number, title, state, labels, assignees	One item redacted by integrity policy
list_pull_requests	array	1	number, title, state, html_url	Flat, no nesting
list_workflows (mcpscripts)	object	2	total_count, per_page, page, workflows[]	Clean pagination envelope
list_code_scanning_alerts	array	1	(empty in this test)	Trivial when no alerts open
list_discussions	object	3	discussions[].category.name, .user.login, pageInfo	Nested category + user + pagination
get_label	object	1	name, color, description, id	Minimal but overhead-heavy
search_code	object	2	items[].name, items[].path, total_count	No snippet/line content
search_users	object	2	items[].login, items[].id, items[].profile_url	No bio/role info

Response Size Analysis

Toolset	Avg Tokens	Tools Tested
repos	6,000	1
discussions	500	1
users	460	1
search	440	1
labels	440	1
actions	170	1
workflow_context	130	1
pull_requests	40	1
issues	45	1
code_security	10	1

Tool-by-Tool Analysis

Tool	Toolset	Tokens	Schema	Rating	Notes
workflow_context	workflow_context	130	markdown	⭐⭐⭐⭐⭐	Injected, no MCP round-trip
list_workflows	actions	170	object	⭐⭐⭐⭐⭐	mcpscripts wrapper, no icon bloat
list_pull_requests	pull_requests	40	array	⭐⭐⭐⭐	Field-selected, tiny and actionable
list_code_scanning_alerts	code_security	10	array	⭐⭐⭐⭐	Empty-state clarity
get_file_contents	repos	6,000	object	⭐⭐⭐	README fetch pulled in huge contributor list
list_issues	issues	45	array	⭐⭐⭐	Integrity filter hid a matching issue
list_discussions	discussions	500	object	⭐⭐⭐	Icon _meta ~3x's the real payload cost
get_label	labels	440	object	⭐⭐⭐	~40 tokens of real data, ~400 of overhead
search_code	search	440	object	⭐⭐⭐	No snippets, requires follow-up read
search_users	users	460	object	⭐⭐⭐	Identity-only, no profile depth

30-Day Trend Summary

Metric	Value
Data Points	10 (day 1 of collection)
Average Daily Tokens	8,235
Average Rating Trend	Not yet available (first data point — will populate as more runs land)

Visualizations

Response Size by Toolset

Usefulness Ratings

Daily Token Trend

Size vs Usefulness

Recommendations

• High-value tools (rating 4-5): list_workflows (via mcpscripts wrapper), list_pull_requests, list_code_scanning_alerts, and the injected workflow_context block — prefer these paths when available.
• Tools needing improvement: raw GitHub MCP responses that repeat the _meta.serverInfo base64 icon block on every call (get_label, search_code, search_users, list_discussions) — this is pure overhead for agentic consumers and dominates token cost for small payloads.
• Context-efficient + high-value: list_workflows (mcpscripts) and workflow_context — both under 200 tokens with a 5/5 rating.
• Context-heavy: get_file_contents on README.md (6,000 tokens) — driven by an auto-generated contributor list, not by the tool itself; prefer range/section reads for large docs when possible.

This is the first day of a 30-day rolling analysis; the trend charts will become meaningful once more daily runs accumulate in /tmp/gh-aw/cache-memory/mcp_analysis.jsonl.

References:

• §32845168085

Generated by 🔍 GitHub MCP Structural Analysis · claude · agent · 119.5 AIC · ⌖ 8.94 AIC · ⊞ 11.5K · ◷

• expires on Aug 26, 2026, 4:10 AM UTC-08:00